Dr.Walter KurzMBA, M.Sc.

Authority, access, budget, the standing to refuse and a route to escalate: what has to be granted for the role to exist.

• Difficulty

Board, statutory officers, process owners, technology and legal: what does not move to the officer.

• Difficulty

Reporting line, access to the accountable body, and the conflict created when the officer also owns delivery.

• Difficulty

The first ninety days: no register, no policy, no budget line, and systems already running.

• Difficulty

Forty people and no lawyer, against a group with three assurance functions.

• Difficulty

Asking for the budget and the people the apparatus needs, with a figure.

• Difficulty

The inventory: system, purpose, owner, supplier, users, affected groups, data, risk class, status.

• Difficulty

Embedded features and supplier-added AI, brought into scope.

• Difficulty

Wiring intake, procurement, security review and change management into the register so it survives a year.

• Difficulty

Turning principles into rules somebody can apply without asking the officer.

• Difficulty

Acceptable use, written in language people read.

• Difficulty

Thresholds, restricted uses, and uses the organisation will not make, translated from oversight intent into an operable taxonomy.

• Difficulty

Who must know what before they may propose, approve, operate or use.

• Difficulty

Who may decide what, at what risk level.

• Difficulty

Data, evaluation, law, security, oversight, benefit and exit as the evidence pack per decision.

• Difficulty

Designing a committee that decides rather than deliberates.

• Difficulty

Triage, fast lanes and a proportionate path for low-risk uses.

• Difficulty

Designing a path people use instead of going around the officer.

• Difficulty

Making a decision reconstructable by somebody who was not in the room.

• Difficulty

Bringing a running system back to the gate before it drifts out of its permission.

• Difficulty

Supplier changes, new features and scope creep after go-live.

• Difficulty

Detection, containment and correction, built before it is needed.

• Difficulty

Running an exercise and acting on what it exposes.

• Difficulty

Failure that arrives as complaints, appeals and quiet workarounds rather than as an alert.

• Difficulty

Who is told, by when, and by whom.

• Difficulty

Holding a position in front of the press and the people affected.

• Difficulty

Designing a route that actually reaches the people who were harmed.

• Difficulty

Channels for affected people and for staff, built before they are needed.

• Difficulty

Stopping something, and surviving having stopped it.

• Difficulty

Leaving a record that protects the organisation and the officer.

• Difficulty

Holding responsibility for a system somebody else runs, including a shared one.

• Difficulty

Supplying each function with what AI obliges it to hold, and noticing when an AI decision lands inside theirs.

• Difficulty

The same, for the systems that answer outward.

• Difficulty

Lines of defence, control testing and corrective action.

• Difficulty

Performance, incidents, drift, benefit and new obligations, reviewed on a rhythm.

• Difficulty

Assembling continuously so the pack already exists when it is asked for.

• Difficulty

Cadence, contents, and how to carry the bad news.

• Difficulty

Making an apparatus stick with people who did not ask for it.

• Difficulty

Regulation, suppliers, technology and incidents elsewhere, feeding the register.

• Difficulty

Putting the pieces together so the whole holds up when it is examined.

• Difficulty

Someone brings the officer a proposal that opens with a technology and a vendor, and never says what is currently going wrong. Everyone in the room nods, because it sounds like progress, and nobody can say what would be different afterwards.

• Difficulty

A proposal says it will "transform customer service" or "automate underwriting". The officer cannot tell what would actually change on a Tuesday, and neither can the people doing the work.

• Difficulty

The officer is asked whether an initiative is worth doing and finds themselves arguing about the technology, because nobody has said where the money or the advantage would actually appear.

• Difficulty

The programme is a year old and someone from finance has asked, reasonably, what it delivered. The pilot report says the system handles the task in four minutes. Nobody wrote down what it took before, nobody kept the old volumes, and the two people who would have known have moved on. The saving may be large. It is now unprovable.

• Difficulty

Two proposals arrive. One saves money on work the whole industry does. The other would let the firm do something competitors cannot. They are presented in the same format, with the same kind of number, and are treated as comparable.

• Difficulty

A vendor demonstration is impressive. The officer is asked whether this would give the firm an edge, and realises the same demonstration is being given to their competitors this week.

• Difficulty

The strategy rests on "our unique data". The officer asks to see it and finds four years of records with the important field filled in half the time, no rights to use it for this purpose, and nothing feeding back from operation.

• Difficulty

An investment case assumes a capability stays scarce for five years. The officer has watched the same capability go from a research demonstration to a checkbox in software they already pay for, twice.

• Difficulty

The business case is approved on a build figure. A year later the bill has grown with usage, nobody budgeted for it, and success has made the finances worse.

• Difficulty

The vendor quote is the number in the paper going to the board. It omits integration, evaluation, monitoring, support, retraining and the change effort, all of which land on the organisation. Two more that are always missing: somebody internally has to own this product, and somebody has to do the assurance and approval work.

• Difficulty

The system is 92 percent accurate and everyone is pleased. Then somebody asks how the reviewer is supposed to know which cases are in the 8 percent. If telling requires redoing the work, the review costs close to the full task rather than eight percent of it, and the saving was never there.

• Difficulty

The efficiency arrives exactly as promised. Within a year every competitor has it, prices have moved, and the customer has the benefit while the firm has the cost.

• Difficulty

The pilot worked. Twelve months later the tool is installed, the benefit case still shows the original figure, and half the team have gone back to the old way of doing it.

• Difficulty

A complete, professional, plausible AI proposal arrives with a decision expected this week. It is the officer's judgement that stands between it and the budget.

• Difficulty

Choosing means foregoing. A plan that rules nothing out has committed to nothing, and this is the sitting where a learner writes the sentence most AI strategies avoid.

• Difficulty

Turning a general wish to use AI into one sentence about how this organisation intends to win with it, and turning scattered pilots into a position.

• Difficulty

Reading the change one level up: what happens to the sector when everyone has the capability, and where that leaves this organisation.

• Difficulty

Turning a list of possible AI uses into a defensible order, on value, feasibility, data readiness, risk and reversibility.

• Difficulty

How much money, across how many bets, over what time: the proportions that turn a ranked list into a portfolio.

• Difficulty

The point where AI stops improving the existing business and starts altering the offering itself.

• Difficulty

Seat, usage, outcome, bundle or tier, and what each does to customer behaviour and to your own economics.

• Difficulty

Keeping an offering profitable when the cost of serving it rises with how much it is used.

• Difficulty

The sourcing decision taken as an argument about what has to be owned to hold a position, rather than as procurement.

• Difficulty

What the supplier owns after signature, and what it would cost to leave.

• Difficulty

Whether timing is decisive for a particular move, and what each posture costs.

• Difficulty

Writing a strategy that survives finding out which of its branches are foreclosed, restricted or expensive to defend.

• Difficulty

Signing a strategy before anyone knows the capability will reach the required quality: staging, thresholds and kill criteria.

• Difficulty

Producing the document itself: what an AI strategy contains and how the parts hold together.

• Difficulty

Holding the position under the four questions a board actually asks: what it costs, what it returns, what a competitor could do about it, and who carries the risk.

• Difficulty

The shift from running projects to running a function: what happens monthly, quarterly and annually, and who is in the room for each.

• Difficulty

Building the path AI requests arrive through, and finding what is already running that nobody asked about.

• Difficulty

The recurring decision: what gets funded this cycle, what continues, what stops, what scales.

• Difficulty

Budgeting a cost that rises with adoption rather than a project that finishes, and handling the variance.

• Difficulty

Measuring what a unit of output actually costs in operation, and watching that number move.

• Difficulty

Changing the work rather than deploying a tool into work that stays the same.

• Difficulty

Carrying the organisation through a change that alters what people do, and what is owed to them while it happens.

• Difficulty

Producing evidence of benefit that a sceptical finance function accepts, rather than a number the programme produced about itself.

• Difficulty

What to do when the benefit did not arrive: continue, pause, stop or scale, and defending the call.

• Difficulty

Drift, degradation, and the supplier model change nobody told you about.

• Difficulty

Running the relationship as a live thing: performance, changes, renegotiation, and the review the contract entitles you to.

• Difficulty

How reversibility is lost quietly, and what has to stay in-house to keep a decision open.

• Difficulty

Deciding what the organisation must be able to do itself, and building for that rather than for everything.

• Difficulty

The standing account to the people entitled to ask: benefit, performance, risk position, and what went wrong.

• Difficulty

Reading operating reality back against the thesis, concluding it is wrong, and forcing the revision.

• Difficulty

Model, training, fine-tuning, prompt, token, context, inference, agent: the vocabulary of a technical meeting, defined well enough to follow one.

• Difficulty

How a model is fitted to data rather than programmed with rules, and what follows from that difference.

• Difficulty

What population a training set represents, and who is missing from it.

• Difficulty

The gap between the outcome an organisation cares about and the measurable thing standing in for it: readmission measured as a billing code, a good hire measured as three years of tenure.

• Difficulty

Trained, fine-tuned, prompted, grounded on your documents: which knob a supplier is turning, and what each one can and cannot fix.

• Difficulty

Predictive, generative, retrieval and acting systems, and what each is genuinely for.

• Difficulty

Problems that do not need a language model, and how to notice them before the project starts.

• Difficulty

Why the same question can produce different answers, and where that rules a use in or out.

• Difficulty

Fluency without a truth check, read as a mechanism rather than as a bug.

• Difficulty

Calibration: how a system's expressed confidence relates, and fails to relate, to whether it is right.

• Difficulty

Context limits, cut-off dates, and the document that was silently truncated.

• Difficulty

Data, objective, deployment context and user behaviour as the four sources, and how to tell which one you are looking at.

• Difficulty

What happens at the edge of the training distribution, and how to predict where a system will fail first.

• Difficulty

Bias as a mechanism rather than an accusation: how the composition of data shapes what a system produces.

• Difficulty

Where ground truth came from, who produced it, under what conditions, and how much to trust it.

• Difficulty

Why a system does not know your documents until it is made to, and what grounding does and does not fix.

• Difficulty

What changes once a model holds tools, credentials and the ability to take actions.

• Difficulty

Images, voice, documents and synthetic media: what these systems do well, where they fail, and where fabricated media becomes a live risk.

• Difficulty

Precision, recall, and what a rare event does to both.

• Difficulty

False positives against false negatives, chosen deliberately and in advance.

• Difficulty

How to read an evaluation claim, and what a benchmark result does not establish about your own use.

• Difficulty

What a demonstration is selected to show, and what turns it into evidence.

• Difficulty

Attack surface as a consequence of the mechanism: a system that takes instructions from text cannot fully separate instruction from content.

• Difficulty

Injection through content the system reads: an email, a document, a web page becoming an instruction.

• Difficulty

Memorisation and extraction: what a model can reveal about what it saw.

• Difficulty

Poisoning, and the provenance of a model or dataset somebody downloaded.

• Difficulty

Where the officer's own competence ends, and how to get an answer they can rely on.

• Difficulty

Drawing the edge of the thing being commissioned: user, task, data, decision, output, handoff.

• Difficulty

From source data through to a person acting on an output, and what sits between.

• Difficulty

Setting the standard a system must meet before anyone builds or sells one.

• Difficulty

Uses whose minimum standard comes from regulation, professional duty or the consequence of error, rather than from the organisation's preference.

• Difficulty

Assembling evaluation material out of the organisation's own reality, including the awkward cases.

• Difficulty

Data held back, and who marks the paper.

• Difficulty

Rubrics, expert review, and what to do when the experts disagree.

• Difficulty

Results broken down by subgroup, language, name, and edge population, rather than averaged.

• Difficulty

Specifying an adversarial test and reading the report it produces.

• Difficulty

Builds the person an organisation holds accountable for its AI: what it does with AI, on what evidence, within what rules, and at what risk. Seven parts and thirteen modules, of which ten are self-learning and three are taught live with a human counterpart, completed with an oral examination.